Training TinyVGG on Flowers17 With Regularization To Improve Accuracy

Import Necessary Packages

from drig.preprocessors import ImageToArrayPreprocessor, UniformAspectPreprocessor
from drig.datum import ImageDatumLoader
from drig.datum_config import Flowers17Config as config
from drig.config import TinyVGGNetImage, ImageCast
from sklearn.preprocessing import LabelBinarizer, LabelEncoder
from sklearn.model_selection import train_test_split
from drig.networks import TinyVGGNet
from keras.optimizers import SGD
from sklearn.metrics import classification_report
from drig.utils import plot_training_metrics, plot_network, \
display_image_data, display_prediction, display_image, \
visualize_network, plot_confusion_mesh, confusion_mesh, random_image, display_prediction, preprocess_image
from keras.preprocessing.image import ImageDataGenerator

Display Sample Image Data

display_image_data(config.DATASET_PATH)

Initialize Preprocessors

config.IMAGE_PREPROCESSING_HEIGHT,config.IMAGE_PREPROCESSING_WIDTH

(64, 64)

uniform_aspect_preproc=UniformAspectPreprocessor(config.IMAGE_PREPROCESSING_HEIGHT, 
                                                 config.IMAGE_PREPROCESSING_WIDTH)
image_to_array_preproc=ImageToArrayPreprocessor()
preprocessors=[uniform_aspect_preproc, image_to_array_preproc]

Load Image Data

data_loader=ImageDatumLoader(preprocessors=preprocessors)
data, labels= data_loader.load(dataset_path=config.DATASET_PATH,
                              class_index=config.CLASS_INDEX,
                              input_cast=ImageCast.RGB_64x64)

Loading Images :100% |❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆❆| 1360 of 1360 Time: 0:00:04
INFO:root:PROCESSED 1360/1360 IMAGES.
INFO:root:MEMORY SIZE OCCUPIED: 65.28 MB

classes=data_loader.classes

Split Train Test Sets

train_x, test_x, train_y, test_y= train_test_split(data, labels, test_size=config.TEST_SIZE, random_state=42)

label_binarizer=LabelBinarizer()
train_y=label_binarizer.fit_transform(train_y)
test_y=label_binarizer.transform(test_y)

Initialize Image Augmenter

augmenter=ImageDataGenerator(rotation_range=30, width_shift_range=0.2, height_shift_range=0.2, 
                             shear_range=0.2, zoom_range=0.2, horizontal_flip=True, fill_mode='nearest')

Compose Network

refiner=SGD(learning_rate=config.ALPHA)
net=TinyVGGNet.compose(height=TinyVGGNetImage.HEIGHT, 
                         width=TinyVGGNetImage.WIDTH,
                         depth=TinyVGGNetImage.DEPTH, 
                         classes=len(classes))
net.compile(loss="categorical_crossentropy", optimizer=refiner, metrics=["accuracy"])

visualize_network(net)

Train Network

Z=net.fit(augmenter.flow(train_x, train_y, batch_size=config.BATCH_SIZE), validation_data=(test_x, test_y),
                     steps_per_epoch=len(train_x)//config.BATCH_SIZE, epochs=config.EPOCHS, verbose=1)

Epoch 1/100
15/15 [==============================] - 3s 160ms/step - loss: 3.6986 - accuracy: 0.1252 - val_loss: 2.8644 - val_accuracy: 0.0500
Epoch 2/100
15/15 [==============================] - 2s 138ms/step - loss: 2.5457 - accuracy: 0.2991 - val_loss: 3.0020 - val_accuracy: 0.0500
Epoch 3/100
15/15 [==============================] - 2s 137ms/step - loss: 2.2982 - accuracy: 0.3358 - val_loss: 3.2049 - val_accuracy: 0.0559
Epoch 4/100
15/15 [==============================] - 2s 137ms/step - loss: 2.0798 - accuracy: 0.3777 - val_loss: 3.2984 - val_accuracy: 0.0529
Epoch 5/100
15/15 [==============================] - 2s 141ms/step - loss: 1.9864 - accuracy: 0.3889 - val_loss: 3.5452 - val_accuracy: 0.0647
Epoch 6/100
15/15 [==============================] - 2s 139ms/step - loss: 1.7956 - accuracy: 0.4493 - val_loss: 3.7944 - val_accuracy: 0.0706
Epoch 7/100
15/15 [==============================] - 2s 136ms/step - loss: 1.7784 - accuracy: 0.4697 - val_loss: 3.8584 - val_accuracy: 0.0882
Epoch 8/100
15/15 [==============================] - 2s 137ms/step - loss: 1.6851 - accuracy: 0.4598 - val_loss: 4.1305 - val_accuracy: 0.0735
Epoch 9/100
15/15 [==============================] - 2s 136ms/step - loss: 1.6045 - accuracy: 0.4698 - val_loss: 4.2278 - val_accuracy: 0.0765
Epoch 10/100
15/15 [==============================] - 2s 136ms/step - loss: 1.6855 - accuracy: 0.4583 - val_loss: 4.2590 - val_accuracy: 0.1265
Epoch 11/100
15/15 [==============================] - 2s 136ms/step - loss: 1.6177 - accuracy: 0.5000 - val_loss: 4.4292 - val_accuracy: 0.1206
Epoch 12/100
15/15 [==============================] - 2s 137ms/step - loss: 1.3569 - accuracy: 0.5509 - val_loss: 4.4959 - val_accuracy: 0.1353
Epoch 13/100
15/15 [==============================] - 2s 136ms/step - loss: 1.3348 - accuracy: 0.5706 - val_loss: 4.8043 - val_accuracy: 0.1147
Epoch 14/100
15/15 [==============================] - 2s 136ms/step - loss: 1.4241 - accuracy: 0.5274 - val_loss: 4.7858 - val_accuracy: 0.1118
Epoch 15/100
15/15 [==============================] - 2s 134ms/step - loss: 1.3732 - accuracy: 0.5398 - val_loss: 5.0138 - val_accuracy: 0.1441
Epoch 16/100
15/15 [==============================] - 2s 137ms/step - loss: 1.2636 - accuracy: 0.5651 - val_loss: 5.2596 - val_accuracy: 0.1412
Epoch 17/100
15/15 [==============================] - 2s 135ms/step - loss: 1.3065 - accuracy: 0.5654 - val_loss: 5.0525 - val_accuracy: 0.1441
Epoch 18/100
15/15 [==============================] - 2s 138ms/step - loss: 1.3604 - accuracy: 0.5385 - val_loss: 4.8987 - val_accuracy: 0.1265
Epoch 19/100
15/15 [==============================] - 2s 137ms/step - loss: 1.2222 - accuracy: 0.6047 - val_loss: 5.1261 - val_accuracy: 0.1441
Epoch 20/100
15/15 [==============================] - 2s 140ms/step - loss: 1.2060 - accuracy: 0.6048 - val_loss: 5.0169 - val_accuracy: 0.1500
Epoch 21/100
15/15 [==============================] - 2s 136ms/step - loss: 1.2057 - accuracy: 0.5923 - val_loss: 4.7557 - val_accuracy: 0.1559
Epoch 22/100
15/15 [==============================] - 2s 137ms/step - loss: 1.1354 - accuracy: 0.6405 - val_loss: 4.2274 - val_accuracy: 0.1706
Epoch 23/100
15/15 [==============================] - 2s 139ms/step - loss: 1.1619 - accuracy: 0.6020 - val_loss: 4.1669 - val_accuracy: 0.2029
Epoch 24/100
15/15 [==============================] - 2s 136ms/step - loss: 1.0954 - accuracy: 0.6240 - val_loss: 3.7753 - val_accuracy: 0.2176
Epoch 25/100
15/15 [==============================] - 2s 139ms/step - loss: 1.0606 - accuracy: 0.6510 - val_loss: 3.6618 - val_accuracy: 0.2235
Epoch 26/100
15/15 [==============================] - 2s 135ms/step - loss: 1.1109 - accuracy: 0.6173 - val_loss: 2.9722 - val_accuracy: 0.2971
Epoch 27/100
15/15 [==============================] - 2s 136ms/step - loss: 1.0990 - accuracy: 0.6387 - val_loss: 2.5667 - val_accuracy: 0.3294
Epoch 28/100
15/15 [==============================] - 2s 134ms/step - loss: 1.0673 - accuracy: 0.6454 - val_loss: 2.6281 - val_accuracy: 0.3118
Epoch 29/100
15/15 [==============================] - 2s 135ms/step - loss: 1.0706 - accuracy: 0.6505 - val_loss: 2.1348 - val_accuracy: 0.3912
Epoch 30/100
15/15 [==============================] - 2s 135ms/step - loss: 1.0790 - accuracy: 0.6275 - val_loss: 2.0965 - val_accuracy: 0.3882
Epoch 31/100
15/15 [==============================] - 2s 137ms/step - loss: 1.0084 - accuracy: 0.6482 - val_loss: 1.7638 - val_accuracy: 0.4265
Epoch 32/100
15/15 [==============================] - 2s 139ms/step - loss: 1.0991 - accuracy: 0.6192 - val_loss: 2.1316 - val_accuracy: 0.3676
Epoch 33/100
15/15 [==============================] - 2s 136ms/step - loss: 0.9908 - accuracy: 0.6588 - val_loss: 1.6760 - val_accuracy: 0.4588
Epoch 34/100
15/15 [==============================] - 2s 134ms/step - loss: 0.9438 - accuracy: 0.6797 - val_loss: 1.5151 - val_accuracy: 0.4882
Epoch 35/100
15/15 [==============================] - 2s 135ms/step - loss: 1.0062 - accuracy: 0.6521 - val_loss: 1.2657 - val_accuracy: 0.5941
Epoch 36/100
15/15 [==============================] - 2s 137ms/step - loss: 0.9277 - accuracy: 0.6943 - val_loss: 1.3006 - val_accuracy: 0.5500
Epoch 37/100
15/15 [==============================] - 2s 138ms/step - loss: 0.9621 - accuracy: 0.6832 - val_loss: 1.4261 - val_accuracy: 0.5647
Epoch 38/100
15/15 [==============================] - 2s 134ms/step - loss: 0.8929 - accuracy: 0.6936 - val_loss: 1.3217 - val_accuracy: 0.5971
Epoch 39/100
15/15 [==============================] - 2s 134ms/step - loss: 1.0414 - accuracy: 0.6441 - val_loss: 1.3023 - val_accuracy: 0.6000
Epoch 40/100
15/15 [==============================] - 2s 135ms/step - loss: 1.0115 - accuracy: 0.6612 - val_loss: 1.1895 - val_accuracy: 0.6000
Epoch 41/100
15/15 [==============================] - 2s 137ms/step - loss: 0.9384 - accuracy: 0.6795 - val_loss: 1.1662 - val_accuracy: 0.6382
Epoch 42/100
15/15 [==============================] - 2s 137ms/step - loss: 0.8632 - accuracy: 0.6898 - val_loss: 1.2668 - val_accuracy: 0.6294
Epoch 43/100
15/15 [==============================] - 2s 135ms/step - loss: 0.9067 - accuracy: 0.7121 - val_loss: 1.1612 - val_accuracy: 0.6382
Epoch 44/100
15/15 [==============================] - 2s 136ms/step - loss: 0.7981 - accuracy: 0.7195 - val_loss: 1.1005 - val_accuracy: 0.6706
Epoch 45/100
15/15 [==============================] - 2s 140ms/step - loss: 0.8774 - accuracy: 0.6957 - val_loss: 1.2868 - val_accuracy: 0.5794
Epoch 46/100
15/15 [==============================] - 2s 137ms/step - loss: 0.8627 - accuracy: 0.7021 - val_loss: 1.1741 - val_accuracy: 0.6206
Epoch 47/100
15/15 [==============================] - 2s 141ms/step - loss: 0.8033 - accuracy: 0.7375 - val_loss: 1.0708 - val_accuracy: 0.6588
Epoch 48/100
15/15 [==============================] - 2s 136ms/step - loss: 0.8936 - accuracy: 0.6799 - val_loss: 1.1406 - val_accuracy: 0.6500
Epoch 49/100
15/15 [==============================] - 2s 136ms/step - loss: 0.7383 - accuracy: 0.7347 - val_loss: 1.1311 - val_accuracy: 0.6706
Epoch 50/100
15/15 [==============================] - 2s 134ms/step - loss: 0.8966 - accuracy: 0.6928 - val_loss: 1.2331 - val_accuracy: 0.6441
Epoch 51/100
15/15 [==============================] - 2s 135ms/step - loss: 0.7856 - accuracy: 0.7301 - val_loss: 1.0999 - val_accuracy: 0.6618
Epoch 52/100
15/15 [==============================] - 2s 135ms/step - loss: 0.7893 - accuracy: 0.7262 - val_loss: 1.2471 - val_accuracy: 0.6588
Epoch 53/100
15/15 [==============================] - 2s 134ms/step - loss: 0.7598 - accuracy: 0.7361 - val_loss: 1.3550 - val_accuracy: 0.6412
Epoch 54/100
15/15 [==============================] - 2s 134ms/step - loss: 0.8876 - accuracy: 0.6890 - val_loss: 1.2510 - val_accuracy: 0.6412
Epoch 55/100
15/15 [==============================] - 2s 134ms/step - loss: 0.8232 - accuracy: 0.7271 - val_loss: 1.2646 - val_accuracy: 0.6500
Epoch 56/100
15/15 [==============================] - 2s 134ms/step - loss: 0.8314 - accuracy: 0.7062 - val_loss: 1.2726 - val_accuracy: 0.6294
Epoch 57/100
15/15 [==============================] - 2s 134ms/step - loss: 0.7398 - accuracy: 0.7580 - val_loss: 1.2262 - val_accuracy: 0.6706
Epoch 58/100
15/15 [==============================] - 2s 134ms/step - loss: 0.8095 - accuracy: 0.7195 - val_loss: 1.3048 - val_accuracy: 0.6176
Epoch 59/100
15/15 [==============================] - 2s 134ms/step - loss: 0.7966 - accuracy: 0.7440 - val_loss: 1.1199 - val_accuracy: 0.6882
Epoch 60/100
15/15 [==============================] - 2s 139ms/step - loss: 0.7299 - accuracy: 0.7583 - val_loss: 1.5429 - val_accuracy: 0.6059
Epoch 61/100
15/15 [==============================] - 2s 136ms/step - loss: 0.7732 - accuracy: 0.7136 - val_loss: 1.1499 - val_accuracy: 0.7000
Epoch 62/100
15/15 [==============================] - 2s 136ms/step - loss: 0.7480 - accuracy: 0.7400 - val_loss: 1.1083 - val_accuracy: 0.6971
Epoch 63/100
15/15 [==============================] - 2s 134ms/step - loss: 0.7851 - accuracy: 0.7385 - val_loss: 1.2622 - val_accuracy: 0.6559
Epoch 64/100
15/15 [==============================] - 2s 136ms/step - loss: 0.7456 - accuracy: 0.7397 - val_loss: 1.1341 - val_accuracy: 0.6941
Epoch 65/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6839 - accuracy: 0.7527 - val_loss: 1.1836 - val_accuracy: 0.6647
Epoch 66/100
15/15 [==============================] - 2s 136ms/step - loss: 0.8147 - accuracy: 0.7250 - val_loss: 1.1616 - val_accuracy: 0.6882
Epoch 67/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6615 - accuracy: 0.7763 - val_loss: 1.1040 - val_accuracy: 0.6824
Epoch 68/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6405 - accuracy: 0.7810 - val_loss: 1.0888 - val_accuracy: 0.6853
Epoch 69/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6791 - accuracy: 0.7360 - val_loss: 1.2243 - val_accuracy: 0.6912
Epoch 70/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6966 - accuracy: 0.7483 - val_loss: 1.3134 - val_accuracy: 0.6147
Epoch 71/100
15/15 [==============================] - 2s 137ms/step - loss: 0.7346 - accuracy: 0.7374 - val_loss: 1.2702 - val_accuracy: 0.6588
Epoch 72/100
15/15 [==============================] - 2s 134ms/step - loss: 0.7069 - accuracy: 0.7475 - val_loss: 1.0883 - val_accuracy: 0.6824
Epoch 73/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6567 - accuracy: 0.7755 - val_loss: 1.1469 - val_accuracy: 0.6794
Epoch 74/100
15/15 [==============================] - 2s 135ms/step - loss: 0.7061 - accuracy: 0.7453 - val_loss: 1.1425 - val_accuracy: 0.6765
Epoch 75/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6640 - accuracy: 0.7631 - val_loss: 1.5011 - val_accuracy: 0.6176
Epoch 76/100
15/15 [==============================] - 2s 136ms/step - loss: 0.7030 - accuracy: 0.7525 - val_loss: 1.1296 - val_accuracy: 0.6706
Epoch 77/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6731 - accuracy: 0.7673 - val_loss: 1.2211 - val_accuracy: 0.6706
Epoch 78/100
15/15 [==============================] - 2s 135ms/step - loss: 0.6610 - accuracy: 0.7620 - val_loss: 1.0747 - val_accuracy: 0.7206
Epoch 79/100
15/15 [==============================] - 2s 134ms/step - loss: 0.7319 - accuracy: 0.7420 - val_loss: 1.1657 - val_accuracy: 0.6941
Epoch 80/100
15/15 [==============================] - 2s 136ms/step - loss: 0.6506 - accuracy: 0.7579 - val_loss: 1.2043 - val_accuracy: 0.6824
Epoch 81/100
15/15 [==============================] - 2s 134ms/step - loss: 0.7053 - accuracy: 0.7584 - val_loss: 1.2890 - val_accuracy: 0.6706
Epoch 82/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6020 - accuracy: 0.7993 - val_loss: 1.1106 - val_accuracy: 0.6794
Epoch 83/100
15/15 [==============================] - 2s 135ms/step - loss: 0.6086 - accuracy: 0.7705 - val_loss: 1.1567 - val_accuracy: 0.6941
Epoch 84/100
15/15 [==============================] - 2s 136ms/step - loss: 0.5959 - accuracy: 0.7844 - val_loss: 1.2846 - val_accuracy: 0.6500
Epoch 85/100
15/15 [==============================] - 2s 135ms/step - loss: 0.5870 - accuracy: 0.7954 - val_loss: 1.2973 - val_accuracy: 0.6471
Epoch 86/100
15/15 [==============================] - 2s 135ms/step - loss: 0.5803 - accuracy: 0.7823 - val_loss: 1.1657 - val_accuracy: 0.6706
Epoch 87/100
15/15 [==============================] - 2s 135ms/step - loss: 0.6760 - accuracy: 0.7655 - val_loss: 1.1525 - val_accuracy: 0.6794
Epoch 88/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6635 - accuracy: 0.7625 - val_loss: 1.0752 - val_accuracy: 0.7382
Epoch 89/100
15/15 [==============================] - 2s 139ms/step - loss: 0.5668 - accuracy: 0.8207 - val_loss: 1.4003 - val_accuracy: 0.6265
Epoch 90/100
15/15 [==============================] - 2s 137ms/step - loss: 0.6519 - accuracy: 0.7806 - val_loss: 1.1828 - val_accuracy: 0.7029
Epoch 91/100
15/15 [==============================] - 2s 139ms/step - loss: 0.6745 - accuracy: 0.7611 - val_loss: 1.1449 - val_accuracy: 0.6824
Epoch 92/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6541 - accuracy: 0.7797 - val_loss: 1.1841 - val_accuracy: 0.7059
Epoch 93/100
15/15 [==============================] - 2s 136ms/step - loss: 0.6181 - accuracy: 0.7764 - val_loss: 1.1351 - val_accuracy: 0.7235
Epoch 94/100
15/15 [==============================] - 2s 135ms/step - loss: 0.5796 - accuracy: 0.7841 - val_loss: 1.2532 - val_accuracy: 0.7059
Epoch 95/100
15/15 [==============================] - 2s 135ms/step - loss: 0.5541 - accuracy: 0.8070 - val_loss: 1.2170 - val_accuracy: 0.7059
Epoch 96/100
15/15 [==============================] - 2s 136ms/step - loss: 0.5536 - accuracy: 0.8295 - val_loss: 1.3549 - val_accuracy: 0.6794
Epoch 97/100
15/15 [==============================] - 2s 136ms/step - loss: 0.5319 - accuracy: 0.7976 - val_loss: 1.1606 - val_accuracy: 0.7118
Epoch 98/100
15/15 [==============================] - 2s 135ms/step - loss: 0.5737 - accuracy: 0.7831 - val_loss: 1.5578 - val_accuracy: 0.6265
Epoch 99/100
15/15 [==============================] - 2s 134ms/step - loss: 0.6258 - accuracy: 0.7888 - val_loss: 1.5045 - val_accuracy: 0.6588
Epoch 100/100
15/15 [==============================] - 2s 134ms/step - loss: 0.5293 - accuracy: 0.7942 - val_loss: 1.3223 - val_accuracy: 0.6882

Run Predictions On Training Set

predictions=net.predict(test_x, batch_size=config.BATCH_SIZE)

print(classification_report(test_y.argmax(axis=1),predictions.argmax(axis=1), target_names=classes))

              precision    recall  f1-score   support

           0       0.36      0.33      0.34        15
           1       0.58      0.95      0.72        20
          10       0.88      0.78      0.82        27
          11       0.43      0.80      0.56        20
          12       0.58      0.94      0.71        16
          13       0.65      0.61      0.63        18
          14       1.00      0.43      0.61        23
          15       0.92      0.55      0.69        20
          16       1.00      0.58      0.73        19
           2       0.83      0.56      0.67        18
           3       1.00      0.77      0.87        22
           4       0.50      0.80      0.62        20
           5       0.94      0.65      0.77        23
           6       0.86      0.90      0.88        21
           7       0.38      0.39      0.38        23
           8       0.76      0.81      0.79        16
           9       1.00      0.84      0.91        19

    accuracy                           0.69       340
   macro avg       0.74      0.69      0.69       340
weighted avg       0.76      0.69      0.69       340

Visualize Network Performance

plot_training_metrics(Z,epochs=config.EPOCHS, inline=True)

mesh, class_name,class_index=confusion_mesh(test_y.argmax(axis=1), 
                                            predictions.argmax(axis=1),
                                           encoded_classes=LabelEncoder().fit_transform(classes),
                                           class_name="16",
                                           classes=classes.tolist())

plot_confusion_mesh(mesh, class_name)

Run Random Prediction

test_image, class_name=random_image(config.DATASET_PATH,class_index=config.CLASS_INDEX)

display_image(image=test_image)

preprocessed_test_image=preprocess_image(preprocessors=preprocessors,image=test_image, for_prediction=True)

test_predict=net.predict(preprocessed_test_image)

display_prediction(test_predict.argmax(axis=1),classes, image=test_image,font_scale=0.7 ,text_thickness=1)